publications

2026

1. preprint

   

   ELF: Embedded Language Flows

   Keya Hu^*, Linlu Qiu^*, Yiyang Lu, Hanhong Zhao, Tianhong Li, Yoon Kim, Jacob Andreas, and Kaiming He

   May 2026

   arXiv Code
2. CVPR

   

   ARC Is a Vision Problem!

   Keya Hu, Ali Cy, Linlu Qiu, Xiaoman Delores Ding, Runqian Wang, Yeyin Eva Zhu, Jacob Andreas, and Kaiming He

   CVPR, Jun 2026

   arXiv Code

2025

1. ICLR

   

   Combining Induction and Transduction for Abstract Reasoning

   Wen-Ding Li^*, Keya Hu^*, Carter Larsen, Yuqing Wu, Simon Alford, Caleb Woo, Spencer M. Dunn, Hao Tang, Michelangelo Naim, Dat Nguyen, Wei-Long Zheng, Zenna Tavares, Yewen Pu＊, and Kevin Ellis＊

   ICLR, Apr 2025

   arXiv PDF

2024

1. NeurIPS

   

   Code Repair with LLMs gives an Exploration-Exploitation Tradeoff

   Hao Tang, Keya Hu, Jin Peng Zhou, Sicheng Zhong, Wei-Long Zheng, Xujie Si, and Kevin Ellis

   NeurIPS, Jul 2024

   arXiv PDF Website
2. EMBC

   

   Contrastive Self-Supervised EEG Representation Learning for Emotion Classification

   Keya Hu, Ren-Jie Dai, Wen-Tao Chen, Hao-Long Yin, Bao-Liang Lu, and Wei-Long Zheng

   EMBC, Jul 2024

   Abs PDF Slides

   Self-supervised learning provides an effective approach to leverage a large amount of unlabeled data. Numerous previous studies have indicated that applying self-supervision to physiological signals can yield better representations of the signals. In the paper, we aim to apply this method to the crucial field of emotion recognition. We perform the experiment with several state-of-the-art contrastive self-supervised methods to explore their effectiveness in pre-training feature encoders on raw electroencephalography (EEG) signals and fine-tuning the pre-trained encoders on the downstream emotion classification tasks. We attempt to vary the proportion of labeled data used during fine-tuning and find that the improvement from self-supervised methods is more pronounced when the proportion of labeled data is small. Additionally, we explore the transferability of the feature encoders pre-trained on various datasets and observe that most self-supervised methods exhibit a certain degree of transferability. Methods that effectively utilize the temporal information in EEG signals show superior stability, accuracy, and transferability.